Long-Term Memory — Animate | AI Wisdom

← Back to topic
🎞 ANIMATE · Mode 3 of 5 · +100 XP— Long-Term Memory
🧬Embed
💾Store
🔍Recall
💉Inject
🎯 Next: Test →
Step 1 / 4
Speed
Embedding
Convert memory content to semantic vector💡 Analogy
Like filing papers under subject headings — the embedding is the subject heading that determines where in the filing cabinet the memory lives.
Each memory is embedded with the same model used for retrieval — this ensures query and memory live in the same vector space.
Console Output
> embed("User prefers dark mode and Python") → Float32Array(1536) [0.042, -0.018, ...]
●long_term_memory.py
EmbeddingL10–12
1from openai import OpenAI
2from pinecone import Pinecone
3import json, hashlib, time
4 
5client = OpenAI()
6pc = Pinecone()
7index = pc.Index("agent-memory")
8 
9MEMORY_NAMESPACE = "user_facts"
10 
11def embed(text: str) -> list[float]:
12    return client.embeddings.create(model="text-embedding-3-small", input=text).data[0].embedding
13 
14def store_memory(content: str, metadata: dict | None = None):
15    vec = embed(content)
16    fact_id = hashlib.md5(content.encode()).hexdigest()[:8]
17    index.upsert(
18        vectors=[{
19            "id": fact_id,
20            "values": vec,
21            "metadata": {"content": content, "timestamp": time.time(), **(metadata or {})}
22        }],
23        namespace=MEMORY_NAMESPACE
24    )
25    return fact_id
26 
27def recall(query: str, top_k: int = 5) -> list[str]:
28    q_vec = embed(query)
29    results = index.query(
30        vector=q_vec, top_k=top_k,
31        include_metadata=True,
32        namespace=MEMORY_NAMESPACE
33    )
34    return [m.metadata["content"] for m in results.matches if m.score > 0.75]
35 
36def chat_with_memory(user_message: str) -> str:
37    relevant_memories = recall(user_message)
38    memory_context = "
39".join(f"- {m}" for m in relevant_memories)
40    messages = [
41        {"role": "system", "content": f"You are a helpful assistant. Use these facts about the user:
42{memory_context}"},
43        {"role": "user", "content": user_message}
44    ]
45    response = client.chat.completions.create(model="gpt-4o", messages=messages)
46    answer = response.choices[0].message.content
47    store_memory(f"User asked: {user_message}. Assistant: {answer[:100]}")
48    return answer
48 lines
Timeline — 4 Steps
Embedding
Convert memory content to semantic vector💡 Analogy
Like filing papers under subject headings — the embedding is the subject heading that determines where in the filing cabinet the memory lives.
Each memory is embedded with the same model used for retrieval — this ensures query and memory live in the same vector space.
Console Output
> embed("User prefers dark mode and Python") → Float32Array(1536) [0.042, -0.018, ...]

Was this page helpful?

Discussion

Convert memory content to semantic vector

💡 Analogy

Like filing papers under subject headings — the embedding is the subject heading that determines where in the filing cabinet the memory lives.

Each memory is embedded with the same model used for retrieval — this ensures query and memory live in the same vector space.

Console Output

> embed("User prefers dark mode and Python") → Float32Array(1536) [0.042, -0.018, ...]

Convert memory content to semantic vector

💡 Analogy

Like filing papers under subject headings — the embedding is the subject heading that determines where in the filing cabinet the memory lives.

Each memory is embedded with the same model used for retrieval — this ensures query and memory live in the same vector space.

Console Output

> embed("User prefers dark mode and Python") → Float32Array(1536) [0.042, -0.018, ...]